Avoid duplicate output in acyclic_breadth_first

[ekaf]

Fix #3244.

Duplicate elimination is a central theme in the literature on the transitive closure of relational operators.

Note that, when the graph is acyclic, duplicate elimination is important only for efficiency, not for termination.

Performance Evaluation of Algorithms for Transitive Closure, Kabler et. al., InformationSystems Vol. 17, No. 5, PP. 415-441, 1992.

With cyclic cases, however, duplicate elimination becomes also crucial for termination, as seen in #2314.

Currently, the closure operation in wordnet.py uses the acyclic_breadth_first() function imported from nltk/util.py, which is supposed to prevent cycles. However, this function presently only detects cyclic children branches, but it does not properly eliminate some duplicates, that are accessible through different parent paths.

Consider for ex. the transitive closure of the hypernyms of 'calamagrostis.n.01', which presently duplicates the nodes that are accessible by two different paths.

With the present PR, the same call does not output duplicates:

from nltk.corpus import wordnet as wn
ss=wn.synset("calamagrostis.n.01")
print(list(ss.closure(lambda s: s.hypernyms())))

[Synset('gramineae.n.01'), Synset('monocot_genus.n.01'), Synset('monocot_family.n.01'), Synset('genus.n.02'), Synset('family.n.06'), Synset('taxonomic_group.n.01'), Synset('biological_group.n.01'), Synset('group.n.01'), Synset('abstraction.n.06'), Synset('entity.n.01')]

[ekaf]

This PR includes #3218, which fixes the problem raised by @rmccampbell, concerning excessive verbosity in the warnings.

[ekaf]

The acyclic_breadth_first() function uses two structures to manage the relation closure: traversed, a set of nodes already seen, and queue, a deque of (child, level) pairs to lookup next.
This picture shows how 'taxonomic_group.n.01', is added to the queue twice, while not yet added to the traversed set. The problem occurs later, when popping duplicates from the queue without checking traversed. So there are two ways of solving the problem, either checking traversed, or preventing duplicates in the queue. This PR implements the first solution, but it might be interesting to consider the second instead.

[ekaf]

Calculating the transitive hypernymic closure of all the 82115 noun synsets in WordNet 3.0 with either solution yields identical closures in both cases, but the solution implemented in this PR is 15% quicker.